Codex 核心概念

model = "gpt-5-codex"
approval_policy = "on-request"
profile = "deep-review"        # 设置默认 profile

[profiles.deep-review]
model = "gpt-5-pro"
model_reasoning_effort = "high"
approval_policy = "never"

[profiles.lightweight]
model = "gpt-4.1"
approval_policy = "untrusted"

model = "gpt-5-codex"                    # 使用的模型
model_provider = "openai"                # 提供商 ID
model_context_window = 200000            # 上下文窗口令牌数
model_max_output_tokens = 8192           # 最大输出令牌数
model_reasoning_effort = "medium"        # minimal | low | medium | high
model_reasoning_summary = "auto"         # auto | concise | detailed | none
model_verbosity = "medium"               # low | medium | high (GPT-5 API)
model_supports_reasoning_summaries = false  # 强制发送推理元数据
model_reasoning_summary_format = "none"  # none | experimental

approval_policy = "untrusted"            # untrusted | on-failure | on-request | never
sandbox_mode = "workspace-write"         # read-only | workspace-write | danger-full-access

[sandbox_workspace_write]
writable_roots = ["/extra/path"]         # 额外可写根目录
network_access = true                    # 允许出站网络访问
exclude_tmpdir_env_var = false           # 排除 $TMPDIR
exclude_slash_tmp = false                # 排除 /tmp

[shell_environment_policy]
inherit = "all"                          # all | core | none
ignore_default_excludes = false          # 保留包含 KEY/SECRET/TOKEN 的变量
exclude = ["CUSTOM_*"]                   # 排除模式(glob)
include_only = ["PATH", "HOME"]          # 白名单(设置后仅保留匹配项)

[shell_environment_policy.set]
MY_VAR = "value"                         # 显式环境变量覆盖

[model_providers.custom]
name = "My Provider"
base_url = "https://api.example.com/v1"
env_key = "CUSTOM_API_KEY"               # 环境变量名
wire_api = "chat"                        # chat | responses
query_params = { "version" = "v1" }
http_headers = { "X-Custom" = "header" }
env_http_headers = { "Auth" = "TOKEN_VAR" }
request_max_retries = 4
stream_max_retries = 5
stream_idle_timeout_ms = 300000

# 项目文档
project_doc_max_bytes = 1048576          # AGENTS.md 最大读取字节数
project_doc_fallback_filenames = [       # AGENTS.md 缺失时的备选文件
  "CODEX.md",
  "PROJECT.md"
]

[projects."/path/to/project"]
trust_level = "trusted"                  # 标记项目为可信

# 历史记录
[history]
persistence = "save-all"                 # save-all | none
max_bytes = 10485760                     # 预留,当前未强制执行

# 通知
notify = ["notify-send", "-t", "5000"]   # 通知命令(接收 JSON)

[tui]
notifications = true                     # 启用 TUI 通知
# 或限制特定事件类型
notifications = ["task_complete", "error"]

# 显示控制
hide_agent_reasoning = false             # 隐藏推理事件
show_raw_agent_reasoning = false         # 显示原始推理内容

# 文件打开器
file_opener = "vscode"                   # vscode | vscode-insiders | windsurf | cursor | none

# OpenTelemetry
[otel]
environment = "dev"                      # 环境标签
exporter = "none"                        # none | otlp-http | otlp-grpc
log_user_prompt = false                  # 导出原始用户提示

# 认证
chatgpt_base_url = "https://custom.url"  # 覆盖 ChatGPT 登录 URL
forced_login_method = "chatgpt"          # chatgpt | api
forced_chatgpt_workspace_id = "uuid"     # 限制到特定工作区

# 实验性配置
instructions = ""                        # 预留,优先使用 AGENTS.md
experimental_instructions_file = "/path/to/instructions.md"  # 替代内置指令

codex \
--model gpt-5-pro \
--config model_reasoning_effort="high" \
--config approval_policy="never" \
--enable web_search_request

# 检查并迁移这些已弃用的配置:
experimental_use_rmcp_client          → features.rmcp_client
experimental_use_exec_command_tool    → features.streamable_shell
experimental_use_unified_exec_tool    → features.unified_exec
include_apply_patch_tool              → features.apply_patch_freeform
tools.web_search                      → features.web_search_request
tools.view_image                      → features.view_image_tool

[features]
rmcp_client = true
streamable_shell = true
unified_exec = true
apply_patch_freeform = true
web_search_request = true
view_image_tool = true